Filtragem online segmentada baseada em redes neurais operando na informação de um calorímetro de altas energias de fina granularidade

Abstract

ATLAS is one of the LHC’s main experiments and aims to investigate the fundamental constituents of matter and their interactions. In the LHC, particles are collided every 25 ns and can reach an energy of up to 14 TeV and generate a large volume of data (70 TB/s). Electrons represent the final states in many decays of interest in physics, being masked by an intense background noise composed of hadronic jets. To deal with the volume of information, ATLAS implements an online filter system to eliminate much of the non-relevant information and preserve the physics of interest. In general, the algorithms used in the simpler first stages of selection should eliminate much of the background noise and allow the algorithms of the later stages, which are more computationally expensive, to be executed only on events close to the physics of interest. Thus, a discriminator based on an ensemble of neural networks, the NeuralRinger , powered by a data compression system that benefits from the energy deposition profile of the particles in the calorimeter is used to make the most efficient decision in the first stage of electron selection in ATLAS. This work aims to present different strategies based on deep learning and information fusion using the signs of calorimeters and track variables, reconstructed in the early stages of the online filtering, in order to improve the selection of electrons, through NeuralRinger , for the beginning of Run 3 . Also, in a long-term scenario, for example Run 4, the development of a framework for the reconstruction and simulation of events based on a generic calorimeter is discussed